Ram blocks with resizeable anonymous allocations under POSIX
v3 cover Ram blocks with resizeable anonymous allocations under POSIX
This series is now based on [1]:
"[PATCH v3 00/13] migrate/ram: Fix resizing RAM blocks while migrating"
We already allow resizable ram blocks for anonymous memory, however, they are not actually resized. All memory is mmaped() R/W, including the memory exceeding the used_length, up to the max_length.
When resizing, effectively only the boundary is moved. Implement actually resizable anonymous allocations and make use of them in resizable ram blocks when possible. Memory exceeding the used_length will be inaccessible. Especially ram block notifiers require care.
Having actually resizable anonymous allocations (via mmap-hackery) allows to reserve a big region in virtual address space and grow the accessible/usable part on demand. Even if "/proc/sys/vm/overcommit_memory" is set to "never" under Linux, huge reservations will succeed. If there is not enough memory when resizing (to populate parts of the reserved region), trying to resize will fail. Only the actually used size is reserved in the OS.
Especially, memory notifiers already handle resizing by first removing the old region, and then re-adding the resized region. prealloc is currently not possible with resizable ram blocks. mlock() should continue to work as is. Resizing is currently rare and must only happen on the start of an incoming migration, or during resets. No code path (except HAX and SEV ram block notifiers) should access memory outside of the usable range - and if we ever find one, that one has to be fixed (I did not identify any).
E.g., virtio-mem [2] wants to reserve big resizable memory regions and grow the usable part on demand. I think this change is worth sending out individually. I did excessive tests of this with virtio-mem (which makes it very easy to trigger resizes), including precopy and postcopy migration.
Accompanied by a bunch of minor fixes and cleanups.
v2 -> v3:
- Rebased on current master/[1].
- "util: vfio-helpers: Factor out and fix processing of existing ram blocks"
-- moved to [1]
- "util: vfio-helpers: Remove Error parameter from qemu_vfio_undo_mapping()" -- Better parch description
- "util/mmap-alloc: Factor out calculation of pagesize to mmap_pagesize()" -- is now "util/mmap-alloc: Factor out calculation of the pagesize for the guard page"
-- Decided to keep special handling for the guard page for now -- Dropped rb's
- "util/mmap-alloc: Prepare for resizeable mmaps"
-- No asserts sizes against the real page size
- "numa: Teach ram block notifiers about resizable ram blocks" -- Split. One part is in [1], the other is now in "numa: Introduce ram_block_notifiers_support_resize()"
- "exec: Ram blocks with resizeable anonymous allocations under POSIX" -- Call qemu_ram_apply_settings() only populated parts. Call it on freshly populated parts when growing.
- Minor changes
v1 -> v2:
- Add "util: vfio-helpers: Fix qemu_vfio_close()"
- Add "util: vfio-helpers: Remove Error parameter from qemu_vfio_undo_mapping()"
- Add "util: vfio-helpers: Factor out removal from qemu_vfio_undo_mapping()"
- "util/mmap-alloc: ..."
-- Minor changes due to review feedback (e.g., assert alignment, return
bool when resizing)
- "util: vfio-helpers: Implement ram_block_resized()"
-- Reserve max_size in the IOVA address space.
-- On resize, undo old mapping and do new mapping. We can later implement
a new ioctl to resize the mapping directly.
- "numa: Teach ram block notifiers about resizable ram blocks" -- Pass size/max_size to ram block notifiers, which makes things easier an cleaner
- "exec: Ram blocks with resizable anonymous allocations under POSIX" -- Adapt to new ram block notifiers
-- Shrink after notifying. Always trigger ram block notifiers on resizes -- Add a safety net that all ram block notifiers registered at runtime
support resizes.
[1] https://lkml.kernel.org/r/[email protected]
[2] https://lore.kernel.org/kvm/[email protected]/
David Hildenbrand (15):
util: vfio-helpers: Fix qemu_vfio_close()
util: vfio-helpers: Remove Error parameter from
qemu_vfio_undo_mapping()
util: vfio-helpers: Factor out removal from qemu_vfio_undo_mapping()
exec: Factor out setting ram settings (madvise ...) into
qemu_ram_apply_settings()
exec: Reuse qemu_ram_apply_settings() in qemu_ram_remap()
exec: Drop "shared" parameter from ram_block_add() util/mmap-alloc: Factor out calculation of the pagesize for the guard
page
util/mmap-alloc: Factor out reserving of a memory region to
mmap_reserve()
util/mmap-alloc: Factor out populating of memory to mmap_populate() util/mmap-alloc: Prepare for resizeable mmaps util/mmap-alloc: Implement resizeable mmaps
util: vfio-helpers: Implement ram_block_resized()
util: oslib: Resizeable anonymous allocations under POSIX
numa: Introduce ram_block_notifiers_support_resize()
exec: Ram blocks with resizeable anonymous allocations under POSIX
exec.c | 99 +++++++++++++++++----- hw/core/numa.c | 19 +++++ include/exec/cpu-common.h | 2 + include/exec/memory.h | 8 ++ include/exec/ramlist.h | 1 + include/qemu/mmap-alloc.h | 21 +++-- include/qemu/osdep.h | 6 +- util/mmap-alloc.c | 167 ++++++++++++++++++++++++-------------- util/oslib-posix.c | 37 ++++++++- util/oslib-win32.c | 14 ++++ util/trace-events | 9 +- util/vfio-helpers.c | 120 ++++++++++++++++++++------- 12 files changed, 380 insertions(+), 123 deletions(-)
Signature · 1 line
2.24.1
PATCH v3 1/15 util: vfio-helpers: Fix qemu_vfio_close()
qemu_vfio_undo_mapping() will decrement the number of mappings and reshuffle the array elements to fit into the reduced size.
Iterating over all elements like this does not work as expected, let's make sure to remove all mappings properly.
--- util/vfio-helpers.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)
util/vfio-helpers.c +2 -4
@@ -695,13 +695,11 @@ static void qemu_vfio_reset(QEMUVFIOState *s)/* Close and free the VFIO resources. */ void qemu_vfio_close (QEMUVFIOState *s) { -int i; -if (!s) {return ; } -for (i =0 ; i < s->nr_mappings; ++i) { -qemu_vfio_undo_mapping (s, &s->mappings[i],NULL ); +while (s->nr_mappings) { +qemu_vfio_undo_mapping (s, &s->mappings[s->nr_mappings -1 ],NULL ); }ram_block_notifier_remove (&s->ram_notifier);qemu_vfio_reset (s);
Signature · 1 line
2.24.1
PATCH v3 4/15 exec: Factor out setting ram settings (madvise ...) into qemu_ram_apply_settings()
In reply to David HildenbrandFactor all settings out into qemu_ram_apply_settings().
For memory_try_enable_merging(), the important bit is that it won't be called with XEN - which is now still the case as new_block->host will remain NULL.
--- exec.c | 28 ++++++++++++++++------------ 1 file changed, 16 insertions(+), 12 deletions(-)
exec.c +16 -12
@@ -2069,6 +2069,21 @@ static int memory_try_enable_merging(void *addr, size_t len)return qemu_madvise (addr, len,QEMU_MADV_MERGEABLE ); } +static void qemu_ram_apply_settings (void *host,size_t length) +{ +memory_try_enable_merging (host, length); +qemu_ram_setup_dump (host, length); +qemu_madvise (host, length,QEMU_MADV_HUGEPAGE ); +/* +* MADV_DONTFORK is also needed by KVM in absence of synchronous MMU +* Configure it unless the machine is a qtest server, in which case KVM is +* not used and it may be forked (eg for fuzzing purposes). +*/ +if (!qtest_enabled ()) { +qemu_madvise (host, length,QEMU_MADV_DONTFORK ); + } +} +/* * Resizing RAM while migrating can result in the migration being canceled. * Care has to be taken if the guest might have already detected the memory.
@@ -2227,7 +2242,6 @@ static void ram_block_add(RAMBlock *new_block, Error **errp, bool shared)qemu_mutex_unlock_ramlist ();return ; } -memory_try_enable_merging (new_block->host, new_block->max_length); } }
@@ -2265,17 +2279,7 @@ static void ram_block_add(RAMBlock *new_block, Error **errp, bool shared) DIRTY_CLIENTS_ALL);if (new_block->host) { -qemu_ram_setup_dump (new_block->host, new_block->max_length); -qemu_madvise (new_block->host, new_block->max_length,QEMU_MADV_HUGEPAGE ); -/* -* MADV_DONTFORK is also needed by KVM in absence of synchronous MMU -* Configure it unless the machine is a qtest server, in which case -* KVM is not used and it may be forked (eg for fuzzing purposes). -*/ -if (!qtest_enabled ()) { -qemu_madvise (new_block->host, new_block->max_length, -QEMU_MADV_DONTFORK ); - } +qemu_ram_apply_settings (new_block->host, new_block->max_length);ram_block_notify_add (new_block->host, new_block->used_length, new_block->max_length); }
Signature · 1 line
2.24.1
PATCH v3 3/15 util: vfio-helpers: Factor out removal from qemu_vfio_undo_mapping()
In reply to David HildenbrandFactor it out and properly use it where applicable. Make qemu_vfio_undo_mapping() look like qemu_vfio_do_mapping(), passing the size and iova, not the mapping.
--- util/vfio-helpers.c | 43 +++++++++++++++++++++++++++---------------- 1 file changed, 27 insertions(+), 16 deletions(-)
util/vfio-helpers.c +27 -16
@@ -518,6 +518,20 @@ static IOVAMapping *qemu_vfio_add_mapping(QEMUVFIOState *s,return insert; } +/** +* Remove the mapping from @s and free it. +*/ +static void qemu_vfio_remove_mapping (QEMUVFIOState *s,IOVAMapping *mapping) +{ +const int index = mapping - s->mappings; + +assert (index >=0 && index < s->nr_mappings); +memmove (mapping, &s->mappings[index +1 ], +sizeof (s->mappings[0 ]) * (s->nr_mappings - index -1 )); + s->nr_mappings--; + s->mappings =g_renew (IOVAMapping, s->mappings, s->nr_mappings); +} +/* Do the DMA mapping with VFIO. */ static int qemu_vfio_do_mapping (QEMUVFIOState *s,void *host,size_t size,uint64_t iova)
@@ -539,29 +553,22 @@ static int qemu_vfio_do_mapping(QEMUVFIOState *s, void *host, size_t size, }/** -* Undo the DMA mapping from @s with VFIO, and remove from mapping list. +* Undo the DMA mapping from @s with VFIO. */ -static void qemu_vfio_undo_mapping (QEMUVFIOState *s,IOVAMapping *mapping) +static void qemu_vfio_undo_mapping (QEMUVFIOState *s,size_t size,uint64_t iova) { -int index;struct vfio_iommu_type1_dma_unmap unmap = { .argsz =sizeof (unmap), .flags =0 , - .iova = mapping->iova, - .size = mapping->size, + .iova = iova, + .size = size, }; - index = mapping - s->mappings; -assert (mapping->size >0 ); -assert (QEMU_IS_ALIGNED (mapping->size, qemu_real_host_page_size)); -assert (index >=0 && index < s->nr_mappings); +assert (size >0 ); +assert (QEMU_IS_ALIGNED (size, qemu_real_host_page_size));if (ioctl (s->container,VFIO_IOMMU_UNMAP_DMA , &unmap)) {error_report ("VFIO_UNMAP_DMA failed: %s" ,strerror (errno)); } -memmove (mapping, &s->mappings[index +1 ], -sizeof (s->mappings[0 ]) * (s->nr_mappings - index -1 )); - s->nr_mappings--; - s->mappings =g_renew (IOVAMapping, s->mappings, s->nr_mappings); }/* Check if the mapping list is (ascending) ordered. */
@@ -621,7 +628,7 @@ int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,assert (qemu_vfio_verify_mappings (s)); ret =qemu_vfio_do_mapping (s, host, size, iova0);if (ret) { -qemu_vfio_undo_mapping (s, mapping); +qemu_vfio_remove_mapping (s, mapping); goto out; } s->low_water_mark += size;
@@ -681,7 +688,8 @@ void qemu_vfio_dma_unmap(QEMUVFIOState *s, void *host)if (!m) { goto out; } -qemu_vfio_undo_mapping (s, m); +qemu_vfio_undo_mapping (s, m->size, m->iova); +qemu_vfio_remove_mapping (s, m); out:qemu_mutex_unlock (&s->lock); }
@@ -698,7 +706,10 @@ void qemu_vfio_close(QEMUVFIOState *s)return ; }while (s->nr_mappings) { -qemu_vfio_undo_mapping (s, &s->mappings[s->nr_mappings -1 ]); +IOVAMapping *m = &s->mappings[s->nr_mappings -1 ]; + +qemu_vfio_undo_mapping (s, m->size, m->iova); +qemu_vfio_remove_mapping (s, m); }ram_block_notifier_remove (&s->ram_notifier);qemu_vfio_reset (s);
Signature · 1 line
2.24.1
PATCH v3 2/15 util: vfio-helpers: Remove Error parameter from qemu_vfio_undo_mapping()
In reply to David HildenbrandEverybody discards the error. Let's error_report() instead so this error doesn't get lost.
This is now the same error handling as in qemu_vfio_do_mapping(). However, we don't report any errors via the return value to the caller. This seems to be one of these "will never happen, but let's better print an error because the caller can't handle it either way" cases.
--- util/vfio-helpers.c | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-)
util/vfio-helpers.c +5 -6
@@ -541,8 +541,7 @@ static int qemu_vfio_do_mapping(QEMUVFIOState *s, void *host, size_t size,/** * Undo the DMA mapping from @s with VFIO, and remove from mapping list. */ -static void qemu_vfio_undo_mapping (QEMUVFIOState *s,IOVAMapping *mapping, -Error **errp) +static void qemu_vfio_undo_mapping (QEMUVFIOState *s,IOVAMapping *mapping) {int index;struct vfio_iommu_type1_dma_unmap unmap = {
@@ -557,7 +556,7 @@ static void qemu_vfio_undo_mapping(QEMUVFIOState *s, IOVAMapping *mapping,assert (QEMU_IS_ALIGNED (mapping->size, qemu_real_host_page_size));assert (index >=0 && index < s->nr_mappings);if (ioctl (s->container,VFIO_IOMMU_UNMAP_DMA , &unmap)) { -error_setg_errno (errp, errno,"VFIO_UNMAP_DMA failed" ); +error_report ("VFIO_UNMAP_DMA failed: %s" ,strerror (errno)); }memmove (mapping, &s->mappings[index +1 ],sizeof (s->mappings[0 ]) * (s->nr_mappings - index -1 ));
@@ -622,7 +621,7 @@ int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,assert (qemu_vfio_verify_mappings (s)); ret =qemu_vfio_do_mapping (s, host, size, iova0);if (ret) { -qemu_vfio_undo_mapping (s, mapping,NULL ); +qemu_vfio_undo_mapping (s, mapping); goto out; } s->low_water_mark += size;
@@ -682,7 +681,7 @@ void qemu_vfio_dma_unmap(QEMUVFIOState *s, void *host)if (!m) { goto out; } -qemu_vfio_undo_mapping (s, m,NULL ); +qemu_vfio_undo_mapping (s, m); out:qemu_mutex_unlock (&s->lock); }
@@ -699,7 +698,7 @@ void qemu_vfio_close(QEMUVFIOState *s)return ; }while (s->nr_mappings) { -qemu_vfio_undo_mapping (s, &s->mappings[s->nr_mappings -1 ],NULL ); +qemu_vfio_undo_mapping (s, &s->mappings[s->nr_mappings -1 ]); }ram_block_notifier_remove (&s->ram_notifier);qemu_vfio_reset (s);
Signature · 1 line
2.24.1
PATCH v3 6/15 exec: Drop "shared" parameter from ram_block_add()
In reply to David HildenbrandProperly store it in the flags of the ram block instead (and the flag even already exists and is used).
E.g., qemu_ram_is_shared() now properly succeeds on all ram blocks that are actually shared.
--- exec.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-)
exec.c +8 -4
@@ -2211,7 +2211,7 @@ static void dirty_memory_extend(ram_addr_t old_ram_size, } } -static void ram_block_add (RAMBlock *new_block,Error **errp,bool shared) +static void ram_block_add (RAMBlock *new_block,Error **errp) {RAMBlock *block;RAMBlock *last_block =NULL ;
@@ -2234,7 +2234,8 @@ static void ram_block_add(RAMBlock *new_block, Error **errp, bool shared) } }else { new_block->host =phys_mem_alloc (new_block->max_length, - &new_block->mr->align, shared); + &new_block->mr->align, +qemu_ram_is_shared (new_block));if (!new_block->host) { error_setg_errno(errp, errno, "cannot set up guest memory '%s'",
@@ -2339,7 +2340,7 @@ RAMBlock *qemu_ram_alloc_from_fd(ram_addr_t size, MemoryRegion *mr,return NULL ; } -ram_block_add (new_block, &local_err, ram_flags &RAM_SHARED ); +ram_block_add (new_block, &local_err);if (local_err) {g_free (new_block);error_propagate (errp, local_err);
@@ -2401,10 +2402,13 @@ RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,if (host) { new_block->flags |=RAM_PREALLOC ; } +if (share) { + new_block->flags |=RAM_SHARED ; + }if (resizeable) { new_block->flags |=RAM_RESIZEABLE ; } -ram_block_add (new_block, &local_err, share); +ram_block_add (new_block, &local_err);if (local_err) {g_free (new_block);error_propagate (errp, local_err);
Signature · 1 line
2.24.1
PATCH v3 5/15 exec: Reuse qemu_ram_apply_settings() in qemu_ram_remap()
In reply to David HildenbrandI don't see why we shouldn't apply all settings to make it look like the surrounding RAM (and enable proper VMA merging).
Note: memory backend settings might have overridden these settings. We would need a callback to let the memory backend fix that up.
--- exec.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
exec.c +1 -2
@@ -2516,8 +2516,7 @@ void qemu_ram_remap(ram_addr_t addr, ram_addr_t length) length, addr);exit (1 ); } -memory_try_enable_merging (vaddr, length); -qemu_ram_setup_dump (vaddr, length); +qemu_ram_apply_settings (vaddr, length); } } }
Signature · 1 line
2.24.1
PATCH v3 7/15 util/mmap-alloc: Factor out calculation of the pagesize for the guard page
In reply to David HildenbrandLet's factor out calculating the size of the guard page and rename the variable to make it clearer that this pagesize only applies to the guard page.
--- util/mmap-alloc.c | 31 ++++++++++++++++--------------- 1 file changed, 16 insertions(+), 15 deletions(-)
util/mmap-alloc.c +16 -15
@@ -82,17 +82,27 @@ size_t qemu_mempath_getpagesize(const char *mem_path)return qemu_real_host_page_size; } +static inline size_t mmap_guard_pagesize (int fd) +{ +#if defined(__powerpc64__) && defined(__linux__) +/* Mappings in the same segment must share the same page size */ +return qemu_fd_getpagesize (fd); +#else +return qemu_real_host_page_size; +#endif +} +void *qemu_ram_mmap (int fd,size_t size,size_t align,bool shared,bool is_pmem) { +const size_t guard_pagesize =mmap_guard_pagesize (fd);int flags;int map_sync_flags =0 ;int guardfd;size_t offset; -size_t pagesize;size_t total;void *guardptr;void *ptr;
@@ -113,8 +123,7 @@ void *qemu_ram_mmap(int fd,* anonymous memory is OK. */ flags =MAP_PRIVATE ; - pagesize =qemu_fd_getpagesize (fd); -if (fd ==-1 || pagesize == qemu_real_host_page_size) { +if (fd ==-1 || guard_pagesize == qemu_real_host_page_size) { guardfd =-1 ; flags |=MAP_ANONYMOUS ; }else {
@@ -123,7 +132,6 @@ void *qemu_ram_mmap(int fd, }#else guardfd =-1 ; - pagesize = qemu_real_host_page_size; flags =MAP_PRIVATE |MAP_ANONYMOUS ;#endif
@@ -135,7 +143,7 @@ void *qemu_ram_mmap(int fd,assert (is_power_of_2 (align));/* Always align to host page size */ -assert (align >= pagesize); +assert (align >= guard_pagesize); flags =MAP_FIXED ; flags |= fd ==-1 ?MAP_ANONYMOUS :0 ;
@@ -189,8 +197,8 @@ void *qemu_ram_mmap(int fd,* a guard page guarding against potential buffer overflows. */ total -= offset; -if (total > size + pagesize) { -munmap (ptr + size + pagesize, total - size - pagesize); +if (total > size + guard_pagesize) { +munmap (ptr + size + guard_pagesize, total - size - guard_pagesize); }return ptr;
@@ -198,15 +206,8 @@ void *qemu_ram_mmap(int fd,void qemu_ram_munmap (int fd,void *ptr,size_t size) { -size_t pagesize; -if (ptr) {/* Unmap both the RAM block and the guard page */ -#if defined(__powerpc64__) && defined(__linux__) - pagesize =qemu_fd_getpagesize (fd); -#else - pagesize = qemu_real_host_page_size; -#endif -munmap (ptr, size + pagesize); +munmap (ptr, size +mmap_guard_pagesize (fd)); } }
Signature · 1 line
2.24.1
PATCH v3 9/15 util/mmap-alloc: Factor out populating of memory to mmap_populate()
In reply to David HildenbrandWe want to populate memory within a reserved memory region. Let's factor that out.
--- util/mmap-alloc.c | 89 +++++++++++++++++++++++++---------------------- 1 file changed, 47 insertions(+), 42 deletions(-)
util/mmap-alloc.c +47 -42
@@ -114,6 +114,50 @@ static void *mmap_reserve(size_t size, int fd)return mmap (0 , size,PROT_NONE , flags, fd,0 ); } +/* +* Populate memory in a reserved region from the given fd (if any). +*/ +static void *mmap_populate (void *ptr,size_t size,int fd,bool shared, +bool is_pmem) +{ +int map_sync_flags =0 ; +int flags =MAP_FIXED ; +void *populated_ptr; + + flags |= fd ==-1 ?MAP_ANONYMOUS :0 ; + flags |= shared ?MAP_SHARED :MAP_PRIVATE ; +if (shared && is_pmem) { + map_sync_flags =MAP_SYNC |MAP_SHARED_VALIDATE ; + } + + populated_ptr =mmap (ptr, size,PROT_READ |PROT_WRITE , + flags | map_sync_flags, fd,0 ); +if (populated_ptr ==MAP_FAILED && map_sync_flags) { +if (errno ==ENOTSUP ) { +char *proc_link =g_strdup_printf ("/proc/self/fd/%d" , fd); +char *file_name =g_malloc0 (PATH_MAX ); +int len =readlink (proc_link, file_name,PATH_MAX -1 ); + +if (len <0 ) { + len =0 ; + } + file_name[len] ='\0' ; +fprintf (stderr,"Warning: requesting persistence across crashes " +"for backend file %s failed. Proceeding without " +"persistence, data might become corrupted in case of host " +"crash.\n" , file_name); +g_free (proc_link); +g_free (file_name); + } +/* +* If mmap failed with MAP_SHARED_VALIDATE | MAP_SYNC, we will try +* again without these flags to handle backwards compatibility. +*/ + populated_ptr =mmap (ptr, size,PROT_READ |PROT_WRITE , flags, fd,0 ); + } +return populated_ptr; +} +static inline size_t mmap_guard_pagesize (int fd) {#if defined(__powerpc64__) && defined(__linux__)
@@ -131,12 +175,8 @@ void *qemu_ram_mmap(int fd, bool is_pmem) {const size_t guard_pagesize =mmap_guard_pagesize (fd); -int flags; -int map_sync_flags =0 ; -size_t offset; -size_t total; -void *guardptr; -void *ptr; +size_t offset, total; +void *ptr, *guardptr;/* * Note: this always allocates at least one extra page of virtual address
@@ -153,44 +193,9 @@ void *qemu_ram_mmap(int fd,/* Always align to host page size */ assert (align >= guard_pagesize); - flags =MAP_FIXED ; - flags |= fd ==-1 ?MAP_ANONYMOUS :0 ; - flags |= shared ?MAP_SHARED :MAP_PRIVATE ; -if (shared && is_pmem) { - map_sync_flags =MAP_SYNC |MAP_SHARED_VALIDATE ; - } - offset =QEMU_ALIGN_UP ((uintptr_t )guardptr, align) - (uintptr_t )guardptr; - ptr =mmap (guardptr + offset, size,PROT_READ |PROT_WRITE , - flags | map_sync_flags, fd,0 ); - -if (ptr ==MAP_FAILED && map_sync_flags) { -if (errno ==ENOTSUP ) { -char *proc_link, *file_name; -int len; - proc_link =g_strdup_printf ("/proc/self/fd/%d" , fd); - file_name =g_malloc0 (PATH_MAX ); - len =readlink (proc_link, file_name,PATH_MAX -1 ); -if (len <0 ) { - len =0 ; - } - file_name[len] ='\0' ; -fprintf (stderr,"Warning: requesting persistence across crashes " -"for backend file %s failed. Proceeding without " -"persistence, data might become corrupted in case of host " -"crash.\n" , file_name); -g_free (proc_link); -g_free (file_name); - } -/* -* if map failed with MAP_SHARED_VALIDATE | MAP_SYNC, -* we will remove these flags to handle compatibility. -*/ - ptr =mmap (guardptr + offset, size,PROT_READ |PROT_WRITE , - flags, fd,0 ); - } - + ptr =mmap_populate (guardptr + offset, size, fd, shared, is_pmem);if (ptr ==MAP_FAILED ) {munmap (guardptr, total);return MAP_FAILED ;
Signature · 1 line
2.24.1
PATCH v3 11/15 util/mmap-alloc: Implement resizeable mmaps
In reply to David HildenbrandImplement resizeable mmaps. For now, the actual resizing is not wired up. Introduce qemu_ram_mmap_resizeable() and qemu_ram_mmap_resize(). Make qemu_ram_mmap() a wrapper of qemu_ram_mmap_resizeable().
--- include/qemu/mmap-alloc.h | 21 +++++++++++-------- util/mmap-alloc.c | 43 ++++++++++++++++++++++++++++----------- 2 files changed, 44 insertions(+), 20 deletions(-)
include/qemu/mmap-alloc.h +13 -8
@@ -7,11 +7,13 @@ size_t qemu_fd_getpagesize(int fd);size_t qemu_mempath_getpagesize (const char *mem_path);/** -* qemu_ram_mmap: mmap the specified file or device. +* qemu_ram_mmap_resizeable: reserve a memory region of @max_size to mmap the +* specified file or device and mmap @size of it. * * Parameters: * @fd: the file or the device to mmap * @size: the number of bytes to be mmaped +* @max_size: the number of bytes to be reserved * @align: if not zero, specify the alignment of the starting mapping address; * otherwise, the alignment in use will be determined by QEMU. * @shared: map has RAM_SHARED flag.
@@ -21,12 +23,15 @@ size_t qemu_mempath_getpagesize(const char *mem_path);* On success, return a pointer to the mapped area. * On failure, return MAP_FAILED. */ -void *qemu_ram_mmap (int fd, -size_t size, -size_t align, -bool shared, -bool is_pmem); - -void qemu_ram_munmap (int fd,void *ptr,size_t size); +void *qemu_ram_mmap_resizeable (int fd,size_t size,size_t max_size, +size_t align,bool shared,bool is_pmem); +bool qemu_ram_mmap_resize (void *ptr,int fd,size_t old_size,size_t new_size, +bool shared,bool is_pmem); +static inline void *qemu_ram_mmap (int fd,size_t size,size_t align, +bool shared,bool is_pmem) +{ +return qemu_ram_mmap_resizeable (fd, size, size, align, shared, is_pmem); +} +void qemu_ram_munmap (int fd,void *ptr,size_t max_size);#endif
util/mmap-alloc.c +31 -12
75 lines not on this page · raw message
Signature · 1 line
2.24.1
PATCH v3 8/15 util/mmap-alloc: Factor out reserving of a memory region to mmap_reserve()
In reply to David HildenbrandWe want to reserve a memory region without actually populating memory. Let's factor that out.
--- util/mmap-alloc.c | 58 +++++++++++++++++++++++++++-------------------- 1 file changed, 33 insertions(+), 25 deletions(-)
util/mmap-alloc.c +33 -25
@@ -82,6 +82,38 @@ size_t qemu_mempath_getpagesize(const char *mem_path)return qemu_real_host_page_size; } +/* +* Reserve a new memory region of the requested size to be used for mapping +* from the given fd (if any). +*/ +static void *mmap_reserve (size_t size,int fd) +{ +int flags =MAP_PRIVATE ; + +#if defined(__powerpc64__) && defined(__linux__) +/* +* On ppc64 mappings in the same segment (aka slice) must share the same +* page size. Since we will be re-allocating part of this segment +* from the supplied fd, we should make sure to use the same page size, to +* this end we mmap the supplied fd. In this case, set MAP_NORESERVE to +* avoid allocating backing store memory. +* We do this unless we are using the system page size, in which case +* anonymous memory is OK. +*/ +if (fd ==-1 ||qemu_fd_getpagesize (fd) == qemu_real_host_page_size) { + fd =-1 ; + flags |=MAP_ANONYMOUS ; + }else { + flags |=MAP_NORESERVE ; + } +#else + fd =-1 ; + flags |=MAP_ANONYMOUS ; +#endif + +return mmap (0 , size,PROT_NONE , flags, fd,0 ); +} +static inline size_t mmap_guard_pagesize (int fd) {#if defined(__powerpc64__) && defined(__linux__)
@@ -101,7 +133,6 @@ void *qemu_ram_mmap(int fd,const size_t guard_pagesize =mmap_guard_pagesize (fd);int flags;int map_sync_flags =0 ; -int guardfd;size_t offset;size_t total;void *guardptr;
@@ -113,30 +144,7 @@ void *qemu_ram_mmap(int fd,*/ total = size + align; -#if defined(__powerpc64__) && defined(__linux__) -/* On ppc64 mappings in the same segment (aka slice) must share the same -* page size. Since we will be re-allocating part of this segment -* from the supplied fd, we should make sure to use the same page size, to -* this end we mmap the supplied fd. In this case, set MAP_NORESERVE to -* avoid allocating backing store memory. -* We do this unless we are using the system page size, in which case -* anonymous memory is OK. -*/ - flags =MAP_PRIVATE ; -if (fd ==-1 || guard_pagesize == qemu_real_host_page_size) { - guardfd =-1 ; - flags |=MAP_ANONYMOUS ; - }else { - guardfd = fd; - flags |=MAP_NORESERVE ; - } -#else - guardfd =-1 ; - flags =MAP_PRIVATE |MAP_ANONYMOUS ; -#endif - - guardptr =mmap (0 , total,PROT_NONE , flags, guardfd,0 ); - + guardptr =mmap_reserve (total, fd);if (guardptr ==MAP_FAILED ) {return MAP_FAILED ; }
Signature · 1 line
2.24.1
PATCH v3 10/15 util/mmap-alloc: Prepare for resizeable mmaps
In reply to David HildenbrandWhen shrinking a mmap we want to re-reserve the already populated area. When growing a memory region, we want to populate starting with a given fd_offset. Prepare by allowing to pass these parameters.
Also, let's make sure we always process full pages, to avoid unmapping/remapping pages that are already in use when growing/shrinking. Add some asserts.
--- util/mmap-alloc.c | 34 +++++++++++++++++++++++----------- 1 file changed, 23 insertions(+), 11 deletions(-)
util/mmap-alloc.c +23 -11
@@ -83,12 +83,12 @@ size_t qemu_mempath_getpagesize(const char *mem_path) }/* -* Reserve a new memory region of the requested size to be used for mapping -* from the given fd (if any). +* Reserve a new memory region of the requested size or re-reserve parts +* of an existing region to be used for mapping from the given fd (if any). */ -static void *mmap_reserve (size_t size,int fd) +static void *mmap_reserve (void *ptr,size_t size,int fd) { -int flags =MAP_PRIVATE ; +int flags =MAP_PRIVATE | (ptr ?MAP_FIXED :0 );#if defined(__powerpc64__) && defined(__linux__)/*
@@ -111,19 +111,23 @@ static void *mmap_reserve(size_t size, int fd) flags |=MAP_ANONYMOUS ;#endif -return mmap (0 , size,PROT_NONE , flags, fd,0 ); +return mmap (ptr, size,PROT_NONE , flags, fd,0 ); }/* * Populate memory in a reserved region from the given fd (if any). */ -static void *mmap_populate (void *ptr,size_t size,int fd,bool shared, -bool is_pmem) +static void *mmap_populate (void *ptr,size_t size,int fd,size_t fd_offset, +bool shared,bool is_pmem) {int map_sync_flags =0 ;int flags =MAP_FIXED ;void *populated_ptr; +if (fd ==-1 ) { + fd_offset =0 ; + } + flags |= fd ==-1 ?MAP_ANONYMOUS :0 ; flags |= shared ?MAP_SHARED :MAP_PRIVATE ;if (shared && is_pmem) {
@@ -131,7 +135,7 @@ static void *mmap_populate(void *ptr, size_t size, int fd, bool shared, } populated_ptr =mmap (ptr, size,PROT_READ |PROT_WRITE , - flags | map_sync_flags, fd,0 ); + flags | map_sync_flags, fd, fd_offset);if (populated_ptr ==MAP_FAILED && map_sync_flags) {if (errno ==ENOTSUP ) {char *proc_link =g_strdup_printf ("/proc/self/fd/%d" , fd);
@@ -153,7 +157,8 @@ static void *mmap_populate(void *ptr, size_t size, int fd, bool shared,* If mmap failed with MAP_SHARED_VALIDATE | MAP_SYNC, we will try * again without these flags to handle backwards compatibility. */ - populated_ptr =mmap (ptr, size,PROT_READ |PROT_WRITE , flags, fd,0 ); + populated_ptr =mmap (ptr, size,PROT_READ |PROT_WRITE , flags, fd, + fd_offset); }return populated_ptr; }
@@ -175,16 +180,19 @@ void *qemu_ram_mmap(int fd, bool is_pmem) {const size_t guard_pagesize =mmap_guard_pagesize (fd); +const size_t pagesize =qemu_fd_getpagesize (fd);size_t offset, total;void *ptr, *guardptr; +g_assert (QEMU_IS_ALIGNED (size, pagesize)); +/* * Note: this always allocates at least one extra page of virtual address * space, even if size is already aligned. */ total = size + align; - guardptr =mmap_reserve (total, fd); + guardptr =mmap_reserve (NULL , total, fd);if (guardptr ==MAP_FAILED ) {return MAP_FAILED ; }
@@ -195,7 +203,7 @@ void *qemu_ram_mmap(int fd, offset =QEMU_ALIGN_UP ((uintptr_t )guardptr, align) - (uintptr_t )guardptr; - ptr =mmap_populate (guardptr + offset, size, fd, shared, is_pmem); + ptr =mmap_populate (guardptr + offset, size, fd,0 , shared, is_pmem);if (ptr ==MAP_FAILED ) {munmap (guardptr, total);return MAP_FAILED ;
@@ -219,6 +227,10 @@ void *qemu_ram_mmap(int fd,void qemu_ram_munmap (int fd,void *ptr,size_t size) { +const size_t pagesize =qemu_fd_getpagesize (fd); + +g_assert (QEMU_IS_ALIGNED (size, pagesize)); +if (ptr) {/* Unmap both the RAM block and the guard page */ munmap (ptr, size +mmap_guard_pagesize (fd));
Signature · 1 line
2.24.1
PATCH v3 13/15 util: oslib: Resizeable anonymous allocations under POSIX
In reply to David HildenbrandIntroduce qemu_anon_ram_alloc_resizeable() and qemu_anon_ram_resize(). Implement them under POSIX and make them return NULL under WIN32.
Under POSIX, we make use of resizeable mmaps. An implementation under WIN32 is theoretically possible AFAIK and can be added later.
In qemu_anon_ram_free(), rename the size parameter to max_size, to make it clearer that we have to use the maximum size when freeing resizeable anonymous allocations.
--- include/qemu/osdep.h | 6 +++++- util/oslib-posix.c | 37 ++++++++++++++++++++++++++++++++++--- util/oslib-win32.c | 14 ++++++++++++++ util/trace-events | 4 +++- 4 files changed, 56 insertions(+), 5 deletions(-)
include/qemu/osdep.h +5 -1
@@ -311,8 +311,12 @@ int qemu_daemon(int nochdir, int noclose);void *qemu_try_memalign (size_t alignment,size_t size);void *qemu_memalign (size_t alignment,size_t size);void *qemu_anon_ram_alloc (size_t size,uint64_t *align,bool shared); +void *qemu_anon_ram_alloc_resizeable (size_t size,size_t max_size, +uint64_t *align,bool shared); +bool qemu_anon_ram_resize (void *ptr,size_t old_size,size_t new_size, +bool shared);void qemu_vfree (void *ptr); -void qemu_anon_ram_free (void *ptr,size_t size); +void qemu_anon_ram_free (void *ptr,size_t max_size);#define QEMU_MADV_INVALID -1
util/oslib-posix.c +34 -3
@@ -223,16 +223,47 @@ void *qemu_anon_ram_alloc(size_t size, uint64_t *alignment, bool shared)return ptr; } +void *qemu_anon_ram_alloc_resizeable (size_t size,size_t max_size, +uint64_t *alignment,bool shared) +{ +size_t align =QEMU_VMALLOC_ALIGN ; +void *ptr =qemu_ram_mmap_resizeable (-1 , size, max_size, align, shared, + false); + +if (ptr ==MAP_FAILED ) { +return NULL ; + } + +if (alignment) { + *alignment = align; + } + +trace_qemu_anon_ram_alloc_resizeable (size, max_size, ptr); +return ptr; +} + +bool qemu_anon_ram_resize (void *ptr,size_t old_size,size_t new_size, +bool shared) +{ +bool resized =qemu_ram_mmap_resize (ptr,-1 , old_size, new_size, shared, + false); + +if (resized) { +trace_qemu_anon_ram_resize (old_size, new_size, ptr); + } +return resized; +} +void qemu_vfree (void *ptr) {trace_qemu_vfree (ptr);free (ptr); } -void qemu_anon_ram_free (void *ptr,size_t size) +void qemu_anon_ram_free (void *ptr,size_t max_size) { -trace_qemu_anon_ram_free (ptr, size); -qemu_ram_munmap (-1 , ptr, size); +trace_qemu_anon_ram_free (ptr, max_size); +qemu_ram_munmap (-1 , ptr, max_size); }void qemu_set_block (int fd)
util/oslib-win32.c +14 -0
@@ -90,6 +90,20 @@ void *qemu_anon_ram_alloc(size_t size, uint64_t *align, bool shared)return ptr; } +void *qemu_anon_ram_alloc_resizeable (size_t size,size_t max_size, +uint64_t *align,bool shared) +{ +/* resizeable ram not implemented yet */ +return NULL ; +} + +bool qemu_anon_ram_resize (void *ptr,size_t old_size,size_t new_size, +bool shared) +{ +/* resizeable ram not implemented yet */ +return false; +} +void qemu_vfree (void *ptr) {trace_qemu_vfree (ptr);
util/trace-events +3 -1
@@ -46,8 +46,10 @@ qemu_co_mutex_unlock_return(void *mutex, void *self) "mutex %p self %p" # oslib-posix.c qemu_memalign(size_t alignment, size_t size, void *ptr) "alignment %zu size %zu ptr %p" qemu_anon_ram_alloc(size_t size, void *ptr) "size %zu ptr %p" +qemu_anon_ram_alloc_resizeable(size_t size, size_t max_size, void *ptr) "size %zu max_size %zu ptr %p" +qemu_anon_ram_resize(size_t old_size, size_t new_size, void *ptr) "old_size %zu new_size %zu ptr %p" qemu_vfree(void *ptr) "ptr %p" -qemu_anon_ram_free(void *ptr, size_t size) "ptr %p size %zu" +qemu_anon_ram_free(void *ptr, size_t max_size) "ptr %p max_size %zu" # hbitmap.c hbitmap_iter_skip_words(const void *hb, void *hbi, uint64_t pos, unsigned long cur) "hb %p hbi %p pos %"PRId64" cur 0x%lx"
Signature · 1 line
2.24.1
PATCH v3 14/15 numa: Introduce ram_block_notifiers_support_resize()
In reply to David HildenbrandWe want to actually use resizeable allocations in resizeable ram blocks (IOW, make everything between used_length and max_length inaccessible) - however, not all ram block notifiers can support that.
Introduce a way to detect if any registered notifier does not support resizes - ram_block_notifiers_support_resize() - which we can later use to fallback to legacy handling if a registered notifier (esp., SEV and HAX) does not support actual resizes.
--- hw/core/numa.c | 12 ++++++++++++ include/exec/ramlist.h | 1 + 2 files changed, 13 insertions(+)
hw/core/numa.c +12 -0
@@ -914,3 +914,15 @@ void ram_block_notify_resize(void *host, size_t old_size, size_t new_size) } } } + +bool ram_block_notifiers_support_resize (void ) +{ +RAMBlockNotifier *notifier; + + QLIST_FOREACH(notifier, &ram_list.ramblock_notifiers, next) { +if (!notifier->ram_block_resized) { +return false; + } + } +return true; +}
include/exec/ramlist.h +1 -0
@@ -79,6 +79,7 @@ void ram_block_notifier_remove(RAMBlockNotifier *n);void ram_block_notify_add (void *host,size_t size,size_t max_size);void ram_block_notify_remove (void *host,size_t size,size_t max_size);void ram_block_notify_resize (void *host,size_t old_size,size_t new_size); +bool ram_block_notifiers_support_resize (void );void ram_block_dump (Monitor *mon);
Signature · 1 line
2.24.1
PATCH v3 12/15 util: vfio-helpers: Implement ram_block_resized()
In reply to David HildenbrandLet's implement ram_block_resized(), allowing resizeable mappings.
For resizeable mappings, we reserve $max_size IOVA address space, but only map $size of it. When resizing, unmap the old part and remap the new part. We'll need a new ioctl to do this atomically (e.g., to resize while the guest is running - not allowed for now).
--- util/trace-events | 5 ++-- util/vfio-helpers.c | 70 ++++++++++++++++++++++++++++++++++++++++----- 2 files changed, 66 insertions(+), 9 deletions(-)
util/trace-events +3 -2
@@ -74,8 +74,9 @@ qemu_mutex_unlock(void *mutex, const char *file, const int line) "released mutex # vfio-helpers.c qemu_vfio_dma_reset_temporary(void *s) "s %p" -qemu_vfio_ram_block_added(void *s, void *p, size_t size) "s %p host %p size 0x%zx" -qemu_vfio_ram_block_removed(void *s, void *p, size_t size) "s %p host %p size 0x%zx" +qemu_vfio_ram_block_added(void *s, void *p, size_t size, size_t max_size) "s %p host %p size 0x%zx max_size 0x%zx" +qemu_vfio_ram_block_removed(void *s, void *p, size_t size, size_t max_size) "s %p host %p size 0x%zx max_size 0x%zx" +qemu_vfio_ram_block_resized(void *s, void *p, size_t old_size, size_t new_sizze) "s %p host %p old_size 0x%zx new_size 0x%zx" qemu_vfio_find_mapping(void *s, void *p) "s %p host %p" qemu_vfio_new_mapping(void *s, void *host, size_t size, int index, uint64_t iova) "s %p host %p size %zu index %d iova 0x%"PRIx64 qemu_vfio_do_mapping(void *s, void *host, size_t size, uint64_t iova) "s %p host %p size %zu iova 0x%"PRIx64
util/vfio-helpers.c +63 -7
134 lines not on this page · raw message
Signature · 1 line
2.24.1
PATCH v3 15/15 exec: Ram blocks with resizeable anonymous allocations under POSIX
In reply to David HildenbrandWe can now make use of resizeable anonymous allocations to implement actually resizeable ram blocks. Resizeable anonymous allocations are not implemented under WIN32 yet and are not available when using alternative allocators. Fall back to the existing handling.
We also have to fallback to the existing handling in case any ram block notifier does not support resizing (esp., AMD SEV, HAX) yet. Remember in RAM_RESIZEABLE_ALLOC if we are using resizeable anonymous allocations.
Try to grow early, as that can easily fail if out of memory. Shrink late and ignore errors (nothing will actually break). Warn only.
The benefit of actually resizeable ram blocks is that e.g., under Linux, only the actual size will be reserved (even if "/proc/sys/vm/overcommit_memory" is set to "never"). Additional memory will be reserved when trying to resize, which allows to have ram blocks that start small but can theoretically grow very large.
Note1: We are not able to create resizeable ram blocks with pre-allocated
memory yet, so prealloc is not affected.
Note2: mlock should work as it used to as os_mlock() does a
mlockall(MCL_CURRENT | MCL_FUTURE), which includes future mappings.
Note3: Nobody should access memory beyond used_length. Memory notifiers
already properly take care of this, only ram block notifiers violate this constraint and, therefore, have to be special-cased. Especially, any ram block notifier that might dynamically register at runtime (e.g., vfio) has to support resizes. Add an assert for that. Both, HAX and SEV register early, so they are fine.
--- exec.c | 64 ++++++++++++++++++++++++++++++++++++--- hw/core/numa.c | 7 +++++ include/exec/cpu-common.h | 2 ++ include/exec/memory.h | 8 +++++ 4 files changed, 76 insertions(+), 5 deletions(-)
exec.c +59 -5
@@ -2001,6 +2001,16 @@ void qemu_ram_unset_migratable(RAMBlock *rb) rb->flags &= ~RAM_MIGRATABLE ; } +bool qemu_ram_is_resizeable (RAMBlock *rb) +{ +return rb->flags &RAM_RESIZEABLE ; +} + +bool qemu_ram_is_resizeable_alloc (RAMBlock *rb) +{ +return rb->flags &RAM_RESIZEABLE_ALLOC ; +} +/* Called with iothread lock held. */ void qemu_ram_set_idstr (RAMBlock *new_block,const char *name,DeviceState *dev) {
@@ -2094,6 +2104,7 @@ static void qemu_ram_apply_settings(void *host, size_t length)*/ int qemu_ram_resize (RAMBlock *block,ram_addr_t newsize,Error **errp) { +const bool shared = block->flags &RAM_SHARED ;const ram_addr_t oldsize = block->used_length;assert (block);
@@ -2104,7 +2115,7 @@ int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)return 0 ; } -if (!(block->flags &RAM_RESIZEABLE )) { +if (!qemu_ram_is_resizeable (block)) { error_setg_errno(errp, EINVAL, "Length mismatch: %s: 0x" RAM_ADDR_FMT " in != 0x" RAM_ADDR_FMT, block->idstr,
@@ -2120,6 +2131,15 @@ int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)return -EINVAL ; } +if (oldsize < newsize &&qemu_ram_is_resizeable_alloc (block)) { +if (!qemu_anon_ram_resize (block->host, oldsize, newsize, shared)) { +error_setg_errno (errp, -ENOMEM ,"Cannot allocate enough memory." ); +return -ENOMEM ; + } +/* apply settings for the newly accessible memory */ +qemu_ram_apply_settings (block->host + oldsize, newsize - oldsize); + } +/* Notify before modifying the ram block and touching the bitmaps. */ if (block->host) {ram_block_notify_resize (block->host, oldsize, newsize);
@@ -2133,6 +2153,16 @@ int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)if (block->resized) { block->resized (block->idstr, newsize, block->host); } + +/* +* Shrinking will only fail in rare scenarios (e.g., maximum number of +* mappings reached), and can be ignored. Warn only. +*/ +if (newsize < oldsize &&qemu_ram_is_resizeable_alloc (block) && + !qemu_anon_ram_resize (block->host, oldsize, newsize, shared)) { +warn_report ("Shrinking memory allocation failed." ); + } +return 0 ; }
@@ -2211,6 +2241,28 @@ static void dirty_memory_extend(ram_addr_t old_ram_size, } } +static void ram_block_alloc_ram (RAMBlock *rb) +{ +const bool shared =qemu_ram_is_shared (rb); + +/* +* If we can, try to allocate actually resizeable ram. Will also fail +* if qemu_anon_ram_alloc_resizeable() is not implemented. +*/ +if (phys_mem_alloc == qemu_anon_ram_alloc && +qemu_ram_is_resizeable (rb) && +ram_block_notifiers_support_resize ()) { + rb->host =qemu_anon_ram_alloc_resizeable (rb->used_length, + rb->max_length, + &rb->mr->align, shared); +if (rb->host) { + rb->flags |=RAM_RESIZEABLE_ALLOC ; +return ; + } + } + rb->host =phys_mem_alloc (rb->max_length, &rb->mr->align, shared); +} +static void ram_block_add (RAMBlock *new_block,Error **errp) {RAMBlock *block;
@@ -2233,9 +2285,7 @@ static void ram_block_add(RAMBlock *new_block, Error **errp)return ; } }else { - new_block->host =phys_mem_alloc (new_block->max_length, - &new_block->mr->align, -qemu_ram_is_shared (new_block)); +ram_block_alloc_ram (new_block);if (!new_block->host) { error_setg_errno(errp, errno, "cannot set up guest memory '%s'",
@@ -2280,7 +2330,11 @@ static void ram_block_add(RAMBlock *new_block, Error **errp) DIRTY_CLIENTS_ALL);if (new_block->host) { -qemu_ram_apply_settings (new_block->host, new_block->max_length); +if (qemu_ram_is_resizeable_alloc (new_block)) { +qemu_ram_apply_settings (new_block->host, new_block->used_length); + }else { +qemu_ram_apply_settings (new_block->host, new_block->max_length); + }ram_block_notify_add (new_block->host, new_block->used_length, new_block->max_length); }
hw/core/numa.c +7 -0
@@ -862,6 +862,13 @@ static int ram_block_notify_add_single(RAMBlock *rb, void *opaque)RAMBlockNotifier *notifier = opaque;if (host) { +/* +* Dynamically adding notifiers that don't support resizes is forbidden +* when dealing with resizeable ram blocks that have actually resizeable +* allocations. +*/ +g_assert (!qemu_ram_is_resizeable_alloc (rb) || + notifier->ram_block_resized); notifier->ram_block_added (notifier, host, size, max_size); }return 0 ;
include/exec/cpu-common.h +2 -0
@@ -66,6 +66,8 @@ void qemu_ram_set_uf_zeroable(RAMBlock *rb);bool qemu_ram_is_migratable (RAMBlock *rb);void qemu_ram_set_migratable (RAMBlock *rb);void qemu_ram_unset_migratable (RAMBlock *rb); +bool qemu_ram_is_resizeable (RAMBlock *rb); +bool qemu_ram_is_resizeable_alloc (RAMBlock *rb);size_t qemu_ram_pagesize (RAMBlock *block);size_t qemu_ram_pagesize_largest (void );
include/exec/memory.h +8 -0
@@ -129,6 +129,14 @@ typedef struct IOMMUNotifier IOMMUNotifier;/* RAM is a persistent kind memory */ #define RAM_PMEM (1 << 5) +/* +* Implies RAM_RESIZEABLE. Memory beyond the used_length is inaccessible +* (esp. initially and after resizing). For such memory blocks, only the +* used_length is reserved in the OS - resizing might fail. Will only be +* used with host OS support and if all ram block notifiers support resizing. +*/ +#define RAM_RESIZEABLE_ALLOC (1 << 6) + static inline void iommu_notifier_init(IOMMUNotifier *n, IOMMUNotify fn, IOMMUNotifierFlag flags, hwaddr start, hwaddr end,
Signature · 1 line
2.24.1
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
In reply to David HildenbrandQuoted 6 lines · David Hildenbrand
Let's implement
ram_block_resized(), allowing resizeable mappings.For resizeable mappings, we reserve $max_size IOVA address space, but only map $size of it. When resizing, unmap the old part and remap the new part. We'll need a new ioctl to do this atomically (e.g., to resize while the guest is running - not allowed for now).
Curious: I think it's true for now because resizing only happens during reboot or destination VM during migration (but before switching). However is that guaranteed too in the future?
[...]
util/vfio-helpers.c int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size, · line 661
Quoted 6 lines
util/vfio-helpers.c +1
-1@@ -631,7 +658,7 @@ int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,qemu_vfio_remove_mapping (s, mapping); goto out; } - s->low_water_mark += size; + s->low_water_mark += max_size;
I think it's fine to only increase the low water mark here, however imo it would be better to also cache the max size in IOVAMapping too, then in resize() we double check new_size <= max_size? It also makes IOVAMapping more self contained.
Thanks,
Signature · 1 line
Peter Xu
Re: [PATCH v3 07/15] util/mmap-alloc: Factor out calculation of the pagesize for the guard page
In reply to David HildenbrandQuoted 11 lines · David Hildenbrand
Let's factor out calculating the size of the guard page and rename the variable to make it clearer that this pagesize only applies to the guard page.
Cc "Michael S. Tsirkin" <[email protected]>Cc Murilo Opsfelder Araujo <[email protected]>Cc Greg Kurz <[email protected]>Cc Eduardo Habkost <[email protected]>Cc "Dr. David Alan Gilbert" <[email protected]>Cc Igor Mammedov <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Signature · 1 line
Peter Xu
Re: [PATCH v3 11/15] util/mmap-alloc: Implement resizeable mmaps
In reply to David HildenbrandQuoted 4 lines · David Hildenbrand
Implement resizeable mmaps. For now, the actual resizing is not wired up. Introduce
qemu_ram_mmap_resizeable()andqemu_ram_mmap_resize(). Makeqemu_ram_mmap()a wrapper ofqemu_ram_mmap_resizeable().
Quoted 9 lines
Cc Richard Henderson <[email protected]>Cc Igor Kotrasinski <[email protected]>Cc Murilo Opsfelder Araujo <[email protected]>Cc "Michael S. Tsirkin" <[email protected]>Cc Greg Kurz <[email protected]>Cc Eduardo Habkost <[email protected]>Cc "Dr. David Alan Gilbert" <[email protected]>Cc Igor Mammedov <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Signature · 1 line
Peter Xu
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
In reply to David HildenbrandQuoted 6 lines · David Hildenbrand
util/vfio-helpers.c +6
-0+static void qemu_vfio_dma_map_resize (QEMUVFIOState *s,void *host, +size_t old_size,size_t new_size) +{ +IOVAMapping *m; +int index =0 ; +
util/vfio-helpers.c out: · line 745
Quoted 10 lines
util/vfio-helpers.c +10
-0+qemu_mutex_lock (&s->lock); + m =qemu_vfio_find_mapping (s, host, &index); +if (!m) { +return ; + } +assert (m->size == old_size); + +/* Note: Not atomic - we need a new ioctl for that. */ +qemu_vfio_undo_mapping (s, m->iova, m->size); +qemu_vfio_do_mapping (s, host, m->iova, new_size);
Another way to ask my previous question 1 (in the other reply): Can we simply map/unmap the extra, while keep the shared untouched?
Thanks,
Quoted 6 lines
util/vfio-helpers.c +6
-0+ + m->size = new_size; +assert (qemu_vfio_verify_mappings (s)); + +qemu_mutex_unlock (&s->lock); +}
Signature · 1 line
Peter Xu
Re: [PATCH v3 13/15] util: oslib: Resizeable anonymous allocations under POSIX
In reply to David HildenbrandQuoted 10 lines · David Hildenbrand
Introduce
qemu_anon_ram_alloc_resizeable()andqemu_anon_ram_resize(). Implement them under POSIX and make them return NULL under WIN32.Under POSIX, we make use of resizeable mmaps. An implementation under WIN32 is theoretically possible AFAIK and can be added later.
In
qemu_anon_ram_free(), rename the size parameter tomax_size, to make it clearer that we have to use the maximum size when freeing resizeable anonymous allocations.
Quoted 8 lines
Cc Richard Henderson <[email protected]>Cc Paolo Bonzini <[email protected]>Cc "Dr. David Alan Gilbert" <[email protected]>Cc Eduardo Habkost <[email protected]>Cc Marcel Apfelbaum <[email protected]>Cc Stefan Weil <[email protected]>Cc Igor Mammedov <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Signature · 1 line
Peter Xu
Re: [PATCH v3 14/15] numa: Introduce ram_block_notifiers_support_resize()
In reply to David HildenbrandQuoted 9 lines · David Hildenbrand
We want to actually use resizeable allocations in resizeable ram blocks (IOW, make everything between
used_lengthandmax_lengthinaccessible) - however, not all ram block notifiers can support that.Introduce a way to detect if any registered notifier does not support resizes -
ram_block_notifiers_support_resize()- which we can later use to fallback to legacy handling if a registered notifier (esp., SEV and HAX) does not support actual resizes.
Quoted 8 lines
Cc Richard Henderson <[email protected]>Cc Paolo Bonzini <[email protected]>Cc "Dr. David Alan Gilbert" <[email protected]>Cc Eduardo Habkost <[email protected]>Cc Marcel Apfelbaum <[email protected]>Cc "Michael S. Tsirkin" <[email protected]>Cc Igor Mammedov <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Signature · 1 line
Peter Xu
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
In reply to Peter XuQuoted 7 lines
Am 28.02.2020 um 20:43 schrieb Peter Xu <[email protected]>:
Quoted 2 lines · David Hildenbrand
Let's implement
ram_block_resized(), allowing resizeable mappings.
Quoted 9 lines
Quoted 4 lines
For resizeable mappings, we reserve $max_size IOVA address space, but only map $size of it. When resizing, unmap the old part and remap the new part. We'll need a new ioctl to do this atomically (e.g., to resize while the guest is running - not allowed for now).
Curious: I think it's true for now because resizing only happens during reboot or destination VM during migration (but before switching). However is that guaranteed too in the future?
E.g., virtio-mem will run mutual exclusive with vfio until vfio won‘t pin all guest memory anymore blindly (iow, is compatible with memory overcommit and discarding of ram blocks).
Quoted 14 lines
[...]
util/vfio-helpers.c int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size, · line 661
Quoted 6 lines
util/vfio-helpers.c +1
-1@@ -631,7 +658,7 @@ int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,qemu_vfio_remove_mapping (s, mapping); goto out; } - s->low_water_mark += size; + s->low_water_mark += max_size;I think it's fine to only increase the low water mark here, however imo it would be better to also cache the max size in IOVAMapping too, then in
resize()we double checknew_size<=max_size? It also makes IOVAMapping more self contained.
I‘ll have a look how much additional code that will imply - if it‘s simple, I‘ll do it.
Thanks!
Quoted 5 lines
Thanks,
Signature · 1 line
Peter Xu
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
In reply to Peter XuQuoted 16 lines
Am 28.02.2020 um 20:55 schrieb Peter Xu <[email protected]>:
Quoted 11 lines · David Hildenbrand
util/vfio-helpers.c +11
-0+static void qemu_vfio_dma_map_resize (QEMUVFIOState *s,void *host, +size_t old_size,size_t new_size) +{ +IOVAMapping *m; +int index =0 ; + +qemu_mutex_lock (&s->lock); + m =qemu_vfio_find_mapping (s, host, &index); +if (!m) { +return ; + }
Quoted 8 lines
util/vfio-helpers.c out: · line 745
Quoted 5 lines
util/vfio-helpers.c +5
-0+assert (m->size == old_size); + +/* Note: Not atomic - we need a new ioctl for that. */ +qemu_vfio_undo_mapping (s, m->iova, m->size); +qemu_vfio_do_mapping (s, host, m->iova, new_size);Another way to ask my previous question 1 (in the other reply): Can we simply map/unmap the extra, while keep the shared untouched?
As far as I understand the kernel implementation, unfortunately no. You might be able to grow (by mapping the delta), but shrinking is not possible AFAIR. And I *think* with many resizes, there could be an issue if I remember correctly.
Thanks!
Quoted 13 lines
Thanks,
Quoted 6 lines
util/vfio-helpers.c +6
-0+ + m->size = new_size; +assert (qemu_vfio_verify_mappings (s)); + +qemu_mutex_unlock (&s->lock); +}Signature · 1 line
Peter Xu
Re: [PATCH v3 15/15] exec: Ram blocks with resizeable anonymous allocations under POSIX
In reply to David HildenbrandOn Thu, Feb 27, 2020 at 11:12:05AM +0100, David Hildenbrand wrote:
[...]
Quoted 6 lines
exec.c +6
-0+static void ram_block_alloc_ram (RAMBlock *rb) +{ +const bool shared =qemu_ram_is_shared (rb); + +/* +* If we can, try to allocate actually resizeable ram. Will also fail
exec.c static void dirty_memory_extend(ram_addr_t old_ram_size, · line 2259
Quoted 10 lines
exec.c +10
-0+* if qemu_anon_ram_alloc_resizeable() is not implemented. +*/ +if (phys_mem_alloc == qemu_anon_ram_alloc && +qemu_ram_is_resizeable (rb) && +ram_block_notifiers_support_resize ()) { + rb->host =qemu_anon_ram_alloc_resizeable (rb->used_length, + rb->max_length, + &rb->mr->align, shared); +if (rb->host) { + rb->flags |=RAM_RESIZEABLE_ALLOC ;
A trivial nit: If it should only be set automatically by the memory code, shall we mark it our somewhere just in case someone passed this in flag explicitly (which, iiuc, is a misuse)? Maybe:
assert(!(rb->flags | RAM_RESIZEABLE_ALLOC));
At the entry of this function? Other than that it looks sane to me:
exec.c static void dirty_memory_extend(ram_addr_t old_ram_size, · line 2264
Quoted 5 lines
exec.c +5
-0+return ; + } + } + rb->host =phys_mem_alloc (rb->max_length, &rb->mr->align, shared); +}
Thanks,
Signature · 1 line
Peter Xu
Peter Xu
Reviewed-byRe: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
In reply to David HildenbrandQuoted 25 lines · David Hildenbrand
Quoted 16 lines
Am 28.02.2020 um 20:55 schrieb Peter Xu <[email protected]>:
Quoted 11 lines · David Hildenbrand
util/vfio-helpers.c +11
-0+static void qemu_vfio_dma_map_resize (QEMUVFIOState *s,void *host, +size_t old_size,size_t new_size) +{ +IOVAMapping *m; +int index =0 ; + +qemu_mutex_lock (&s->lock); + m =qemu_vfio_find_mapping (s, host, &index); +if (!m) { +return ; + }Quoted 8 lines
util/vfio-helpers.c out: · line 745
Quoted 5 lines
util/vfio-helpers.c +5
-0+assert (m->size == old_size); + +/* Note: Not atomic - we need a new ioctl for that. */ +qemu_vfio_undo_mapping (s, m->iova, m->size); +qemu_vfio_do_mapping (s, host, m->iova, new_size);Another way to ask my previous question 1 (in the other reply): Can we simply map/unmap the extra, while keep the shared untouched?
Quoted 1 line
As far as I understand the kernel implementation, unfortunately no. You might be able to grow (by mapping the delta), but shrinking is not possible AFAIR. And I *think* with many resizes, there could be an issue if I remember correctly.
Ah right (and I think the same thing will happen to vfio-pci during memory_region_set_size()). Then please double confirm that virtio-mem is disabled for both vfio-helper users and vfio-pci. Also, would you mind comment above the unmap+map with more information? E.g.:
For now, we must unmap the whole mapped range first and remap with the new size. The reason is that VFIO_IOMMU_UNMAP_DMA might fail with partial unmap of previous mappings, so even we can add extra new mappings to extend the old range, however we still won't able to always success on shrinking memories. The side effect is that it's never safe to do resizing during VM execution.
Or something better. With an enriched comment, I think it's fine:
Thanks,
Signature · 1 line
Peter Xu
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
Quoted 30 lines
Am 28.02.2020 um 21:49 schrieb Peter Xu <[email protected]>:
Quoted 25 lines · David Hildenbrand
Quoted 24 lines
Quoted 3 lines
Am 28.02.2020 um 20:55 schrieb Peter Xu <[email protected]>:
Quoted 6 lines · David Hildenbrand
util/vfio-helpers.c +6
-0+static void qemu_vfio_dma_map_resize (QEMUVFIOState *s,void *host, +size_t old_size,size_t new_size) +{ +IOVAMapping *m; +int index =0 ; +util/vfio-helpers.c out: · line 745
Quoted 10 lines
util/vfio-helpers.c +10
-0+qemu_mutex_lock (&s->lock); + m =qemu_vfio_find_mapping (s, host, &index); +if (!m) { +return ; + } +assert (m->size == old_size); + +/* Note: Not atomic - we need a new ioctl for that. */ +qemu_vfio_undo_mapping (s, m->iova, m->size); +qemu_vfio_do_mapping (s, host, m->iova, new_size);Another way to ask my previous question 1 (in the other reply): Can we simply map/unmap the extra, while keep the shared untouched?
Quoted 5 lines
Quoted 1 line
As far as I understand the kernel implementation, unfortunately no. You might be able to grow (by mapping the delta), but shrinking is not possible AFAIR. And I *think* with many resizes, there could be an issue if I remember correctly.
Ah right (and I think the same thing will happen to vfio-pci during
memory_region_set_size()). Then please double confirm that virtio-mem is disabled for both vfio-helper users and vfio-pci. Also, would you
Yes, will double check, should have taken care of both.
Quoted 2 lines
mind comment above the unmap+map with more information? E.g.:
Sure, will add - thanks!
Quoted 11 lines
For now, we must unmap the whole mapped range first and remap with the new size. The reason is that
VFIO_IOMMU_UNMAP_DMAmight fail with partial unmap of previous mappings, so even we can add extra new mappings to extend the old range, however we still won't able to always success on shrinking memories. The side effect is that it's never safe to do resizing during VM execution.Or something better. With an enriched comment, I think it's fine:
Reviewed-by Peter Xu <[email protected]>
Thanks Peter! Have a nice weekend!
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
In reply to David HildenbrandQuoted 1 line · David Hildenbrand
[...]
Quoted 13 lines
Quoted 12 lines
util/vfio-helpers.c int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size, · line 661
Quoted 6 lines
util/vfio-helpers.c +1
-1@@ -631,7 +658,7 @@ int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,qemu_vfio_remove_mapping (s, mapping); goto out; } - s->low_water_mark += size; + s->low_water_mark += max_size;I think it's fine to only increase the low water mark here, however imo it would be better to also cache the max size in IOVAMapping too, then in
resize()we double checknew_size<=max_size? It also makes IOVAMapping more self contained.
Quoted 1 line
I‘ll have a look how much additional code that will imply - if it‘s simple, I‘ll do it.
AFAICT it'll be as simple as introducing IOVAMapping.max_size, then pass max_size into qemu_vfio_add_mapping(). Thanks,
Signature · 1 line
Peter Xu
Re: [PATCH v3 12/15] util: vfio-helpers: Implement ram_block_resized()
Quoted 18 lines · Peter Xu
Quoted 1 line · David Hildenbrand
[...]
Quoted 13 lines
Quoted 12 lines
util/vfio-helpers.c int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size, · line 661
Quoted 6 lines
util/vfio-helpers.c +1
-1@@ -631,7 +658,7 @@ int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,qemu_vfio_remove_mapping (s, mapping); goto out; } - s->low_water_mark += size; + s->low_water_mark += max_size;I think it's fine to only increase the low water mark here, however imo it would be better to also cache the max size in IOVAMapping too, then in
resize()we double checknew_size<=max_size? It also makes IOVAMapping more self contained.
Quoted 5 lines
Quoted 1 line
I‘ll have a look how much additional code that will imply - if it‘s simple, I‘ll do it.
AFAICT it'll be as simple as introducing IOVAMapping.max_size, then pass
max_sizeintoqemu_vfio_add_mapping(). Thanks,
Yeah, was answering from my mobile without code at hand :) added! Thanks!
Signature · 3 lines
Thanks, David / dhildenb
Re: [PATCH v3 15/15] exec: Ram blocks with resizeable anonymous allocations under POSIX
In reply to Peter XuQuoted 25 lines · Peter Xu
On Thu, Feb 27, 2020 at 11:12:05AM +0100, David Hildenbrand wrote:
[...]
Quoted 6 lines
exec.c +6
-0+static void ram_block_alloc_ram (RAMBlock *rb) +{ +const bool shared =qemu_ram_is_shared (rb); + +/* +* If we can, try to allocate actually resizeable ram. Will also fail exec.c static void dirty_memory_extend(ram_addr_t old_ram_size, · line 2259
Quoted 10 lines
exec.c +10
-0+* if qemu_anon_ram_alloc_resizeable() is not implemented. +*/ +if (phys_mem_alloc == qemu_anon_ram_alloc && +qemu_ram_is_resizeable (rb) && +ram_block_notifiers_support_resize ()) { + rb->host =qemu_anon_ram_alloc_resizeable (rb->used_length, + rb->max_length, + &rb->mr->align, shared); +if (rb->host) { + rb->flags |=RAM_RESIZEABLE_ALLOC ;A trivial nit: If it should only be set automatically by the memory code, shall we mark it our somewhere just in case someone passed this in flag explicitly (which, iiuc, is a misuse)? Maybe:
Quoted 3 lines
assert(!(rb->flags | RAM_RESIZEABLE_ALLOC));At the entry of this function? Other than that it looks sane to me:
Excellent idea, thanks!
Signature · 3 lines
Thanks, David / dhildenb
Re: [PATCH v3 09/15] util/mmap-alloc: Factor out populating of memory to mmap_populate()
In reply to David HildenbrandQuoted 14 lines · David Hildenbrand
We want to populate memory within a reserved memory region. Let's factor that out.
Reviewed-by Richard Henderson <[email protected]>Reviewed-by Peter Xu <[email protected]>Cc Igor Kotrasinski <[email protected]>Cc "Michael S. Tsirkin" <[email protected]>Cc Greg Kurz <[email protected]>Cc Murilo Opsfelder Araujo <[email protected]>Cc Eduardo Habkost <[email protected]>Cc "Dr. David Alan Gilbert" <[email protected]>Cc Igor Mammedov <[email protected]>Signed-off-by David Hildenbrand <[email protected]>
Thought a bit longer about the terminology (as brought up by Peter). Populate might indeed be confusing, because we are not actually populating all memory (e.g., MAP_POPULATE). We enable/activate the mapping to be accessible.
Hmmm
mmap_enable() ?
mmap_activate() ?
I think I like the latter. Alternatives?
Signature · 3 lines
Thanks, David / dhildenb